On February 12, 2026, MiniMax released M2.5 and M2.5-Lightning and put the weights on Hugging Face the same day. The positioning fits in one sentence: frontier-adjacent capability at a fraction of the price. The model scores 80.2% on SWE-Bench Verified and finishes tasks in an average of 22.8 minutes — level with Claude Opus 4.6’s 22.9 minutes — at roughly a tenth of the cost.
The timing lands in the middle of China’s model release blitz. Zhipu had open-sourced GLM-5 the day before (see our earlier coverage), and VentureBeat led its M2.5 story on near-state-of-the-art performance at a twentieth of Claude Opus 4.6’s cost. Cheap plus open is no longer one company’s strategy; it is the market’s default setting.
Two Variants: Same Brains, Double the Speed
M2.5 and M2.5-Lightning are identical in capability and differ only in speed and price. M2.5 runs at 50 tokens per second, at $0.15 per million input tokens and $1.20 per million output tokens. Lightning runs at 100 tokens per second — which MiniMax claims is twice most frontier models — at $0.30 input and $2.40 output, and both variants support caching. The announcement converts this into numbers you can feel: an hour of continuous Lightning costs $1, M2.5 just $0.30, and four always-on instances for a full year total $10,000. The model card’s own phrase for this is “intelligence too cheap to meter.” For context, a single SWE-Bench Verified task consumes about 3.52 million tokens on average — a workload that would be eye-watering on frontier-model pricing and routine here.
80.2% SWE-Bench and Cross-Harness Consistency
Beyond the headline 80.2% on SWE-Bench Verified, the more persuasive evidence is cross-harness testing: on Droid, M2.5 scores 79.7 against Opus 4.6’s 78.9; on OpenCode it is 76.1 to 75.9. A single benchmark can be gamed; consistency across environments — including harnesses the model was not tuned for — is what agent engineers actually buy. Other reported numbers: 51.3% on Multi-SWE-Bench, 76.3% on BrowseComp with context management, roughly 20% fewer search and tool-calling rounds than M2.1, and a VIBE-Pro result MiniMax describes as on par with Opus 4.5. For office work, MiniMax reports a 59.0% average win rate on its GDPval-MM evaluation, tied to Office Skills for Word, PowerPoint, and Excel inside MiniMax Agent plus more than ten thousand user-built Experts.
Forge, CISPO, and Agentic RL at Scale
The training recipe is a snapshot of the shared playbook across this wave of Chinese models: agentic reinforcement learning at scale. MiniMax ran RL over more than 200,000 real-world coding environments spanning 10-plus languages and full-stack work across Web, Android, iOS, and Windows. Two in-house components do the heavy lifting: Forge, an agent-native RL framework that decouples agents from the training-inference engine and uses tree-structured sample merging with async scheduling for a roughly 40x training speedup; and CISPO, an algorithm for stabilizing large-scale MoE training. Reward design mixes process rewards with trajectory-level task-completion-time optimization — the model is forced to trade off being smart against being fast, which is where that 22.8-minute average comes from.
Open Weights and Self-Hosting
The weights ship on Hugging Face under a modified MIT license: 229B parameters, stored mostly in FP8 quantization, roughly 230GB to download. Deployment paths are the usual suspects — SGLang, vLLM, Transformers, and KTransformers. Set that against MiniMax’s internal numbers — M2.5 now autonomously handles 30% of tasks inside the company and accounts for 80% of newly committed code — and the bar for running your own “primary coding model” has, for the first time, dropped to a multi-GPU server plus an open weights download.
What It Means for Developers
Three takeaways. First, the marginal cost of coding agents is collapsing: output pricing sits at a tenth to a twentieth of Opus-class models, which turns long-running agents from a cost-control problem into a cost-you-stop-measuring non-problem. Second, open weights change procurement: any team that can self-host 229B now has an anchor for every API negotiation. Third, speed is now a spec: Lightning’s 100 tokens per second brings interactive agents close to the feel of local tooling, which is both a new constraint and a differentiator for IDE and terminal agent UX.
Sources
- MiniMax M2.5: Built for Real-World Productivity — MiniMax
- MiniMax-M2.5 model card — Hugging Face
- MiniMax’s new open M2.5 and M2.5-Lightning near state-of-the-art — VentureBeat
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.